Probabilistic reasoning · AlphaGo · Robotics
I build machines that reason and act — from probabilistic inference at web scale, through AlphaGo, to a new venture bringing AlphaGo-style reasoning to robots.
Chair of Machine Learning, UCL
I am a machine-learning researcher who has spent a career teaching machines to reason under uncertainty and to act in the world. I trained as a physicist — in Hamburg, at Imperial College London, and in Berlin, where I earned my PhD in machine learning in 2001. At Microsoft Research from 2003 I co-founded the Online Services and Advertising group, where probabilistic reasoning shipped at scale: TrueSkill, the Bayesian rating system behind matchmaking on Xbox Live, and AdPredictor, the click-through model behind Bing — both built on approximate message passing over factor graphs.
At DeepMind I returned to my first love — creating intelligent systems — and had the fortune to contribute to AlphaGo, the first program to defeat a professional at the full-sized game of Go, and to its successors AlphaGo Zero, AlphaZero and MuZero. There, learned reasoning met deep search: a policy that suggests, a value that judges, a tree that looks ahead.
Much of my work since has been on multi-agent systems — from Capture-the-Flag agents to simulated humanoid football — where intelligence is embodied, cooperative, and learned through self-play. After leading machine learning for cellular rejuvenation at Altos Labs and a spell on Google DeepMind's Post-AGI team, those threads now converge in a new venture: bringing AlphaGo-style reasoning to robotics, so that machines can plan and act under real-world uncertainty. Alongside this I hold the Chair of Machine Learning at UCL.
Long before deep networks, I worked on machines that reason with uncertainty — Bayesian models that hold beliefs, update them from evidence, and scale to the web through approximate message passing on factor graphs. Ranking players, predicting clicks, recommending, and — a decade before AlphaGo — predicting the next move in Go.
The Bayesian move-prediction research didn't just stay in the lab — it shipped as a game. The Path of Go, an Xbox Live Arcade title from Microsoft Research, let players take on an AI opponent whose moves were guided by the same pattern-ranking models. Built with my wonderful collaborators — David Stern, Joaquín Quiñonero Candela and Ralf Herbrich, all in the video above. Probabilistic reasoning, playable on a console.
Full list on Google Scholar.
Some of the most thrilling work of my life. I was fortunate to be on the team that built AlphaGo — the first program to beat a professional Go player, a decade before anyone expected — and its successors AlphaGo Zero, AlphaZero and MuZero: learning tabula rasa, generalising to chess and shogi, and finally mastering games without even being told the rules. I still relive those days in London and Seoul.
Much of the world's complexity is agents interacting — cells in a body, players in a team, robots in a warehouse. In games we can build embodied, cooperative agents and watch intelligence emerge from self-play. These are the testbeds on the road from AlphaGo to robots that reason and act in the physical world.
A population of reinforcement-learning agents, trained from only pixels and game points, reached human level at Quake III Arena Capture-the-Flag. Through population-based training across thousands of parallel matches on randomly generated maps, the agents learned teamwork, roles, and even their own internal rewards.
Teams of simulated humanoids learn football end to end — from low-level motor control (running, turning, kicking) up to coordinated team play. Division of labour and off-ball movement emerge from self-play: a study of how embodied skill and cooperation can be learned together, a step toward robots that move and decide.
A career rarely runs in a straight line. A few other threads I care about — the road from AGI to superintelligence, cooperation between AIs, reversing ageing in cells, and what our digital traces quietly reveal about us.
A few things I built for the joy of it — small interactive demos exploring ideas in learning, perception, and physics. Best viewed on a desktop browser.
Play a MIDI file and watch the music become light: each note drives a vertical grating whose spatial frequency tracks its pitch, and overlapping notes bloom into moiré patterns.
open demo ↗
A tennis match-probability calculator. Give it a player's chance of winning a point on serve and it computes — via absorbing Markov chains — the exact probability of winning from any score.
open demo ↗
A WebGPU sandbox where coupled oscillators on a diamond lattice self-organise — phases synchronising while bond couplings drift toward criticality. Order emerging from local rules.
open demo ↗
Substack · explaining the world from first principles
Essays on the mental models that cut across disciplines — simple, general, useful patterns like filters, leverage, and order — explored through examples, games, puzzles, and paradoxes. Grasp them once and you start to see them everywhere.
Read on Substack ↗Places that have shaped me — from a Hamburg schoolroom and a tall ship, through universities and research labs, to the boards I serve on today.
Gelehrtenschule des Johanneums, Hamburg
Gorch Fock · German Navy
Georgia Southern University
University of Hamburg
Imperial College London
Technical University of Berlin
RIKEN, Tokyo
ETH Zurich
Royal Holloway, University of London
Microsoft Research Cambridge
Altos Labs
UCL
Google DeepMind
Partnership on AI
Cooperative AI Foundation
London Institute for Mathematical SciencesReach me on any of these, or read the papers above.